Training Flow Matching: The Role of Weighting and Parameterization

TL;DR

This paper systematically analyzes how loss weighting and output parameterization affect training in denoising-based generative models, providing practical insights without proposing new methods.

Contribution

It offers a comprehensive study on the interaction of training choices with data geometry, architecture, and dataset size in flow matching models.

Findings

Training choices significantly influence denoising accuracy and generative quality.

Different parameterizations interact with data manifold complexity and model architecture.

Insights help guide practical design decisions in training flow matching models.

Abstract

We study the training objectives of denoising-based generative models, with a particular focus on loss weighting and output parameterization, including noise-, clean image-, and velocity-based formulations. Through a systematic numerical study, we analyze how these training choices interact with the intrinsic dimensionality of the data manifold, model architecture, and dataset size. Our experiments span synthetic datasets with controlled geometry as well as image data, and compare training objectives using quantitative metrics for denoising accuracy (PSNR across noise levels) and generative quality (FID). Rather than proposing a new method, our goal is to disentangle the various factors that matter when training a flow matching model, in order to provide practical insights on design choices.